Best AI Inference Accelerator AI Tools & Models - Premium AI Inference Accelerator News

AI News

South Korean FuriosaAI Unveils Third-Generation AI Inference Accelerator: 32x Improvement in Computing Power and Memory Bandwidth

South Korean AI chip startup FuriosaAI has unveiled its third-generation AI inference accelerator: developed in collaboration with Broadcom, it includes 2 2nm compute dies, 1 I/O die, and 12 HBM4(E) stacks of 48GB each, supports PCIe Gen7, and features full SUE connectivity within a rack. The size is 8 times that of the second generation, with 32 PFLOPS of computing power, which is 32 times that of the second generation.

10.1k 10 minutes ago

Xiaomi Releases Xuanjie O100 Chip, With the Highest Edge-side Large Model Inference Speed of 330 TPS

Xiaomi unveiled its first on-device LLM AI accelerator chip, Xuanjie O100. Built on 6nm wafer-level vertical stacking and hybrid bonding with 1.4μm pitch, it delivers 1.22TB/s bandwidth—16x that of flagship phones—and up to 330 TPS inference, boosting on-device LLM throughput and reasoning.....

34.4k 1 hours ago

New Breakthrough in Domestic Computing Power! Mooley × Silicon Flow Achieve Efficient Inference of DeepSeek V3 671B on MTT S5000, Single Card Performance Approaching International Top Standards

Domestic AI chips and large models have achieved significant progress in collaborative optimization. Moore Threads and Silicon-based Flow successfully adapted the 671B-parameter DeepSeek V3 model to the domestic GPU MTT S5000. Using FP8 low-precision inference, they achieved over 4000 tokens/sec prefill and over 1000 tokens/sec decoding throughput, nearing international high-end AI accelerator performance.....

21.4k 12 hours ago

ASUS Launches Plug-and-Play AI Acceleration Card UGen300: 40 TOPS Computing Power in USB, Turning Ordinary Computers into AI Terminals

ASUS launches the UGen300 AI accelerator card with a USB interface, offering 40 TOPS AI inference via USB 3.1 Gen2 without needing PCIe slots. It features the Hailo-10H NPU chip, 2.5W power consumption, 8GB onboard memory, and supports large pre-trained models.....

36.2k 7 hours ago

AI Products

Procyon AI Image Generation Benchmark

A benchmarking tool used to measure the AI accelerator inference performance of devices.

AI model

11.7k

MTIA

Meta Training and Inference Accelerator (MTIA) is Meta's self-developed next-generation AI training and inference acceleration chip. Compared to the previous generation, MTIA v2 has significantly improved computing power and memory bandwidth, enabling more efficient support for Meta's sorting and recommendation model applications. This is a significant achievement in Meta's continuous investment in self-developed hardware infrastructure, which will bring new capabilities to Meta's AI products and services.

Hardware infrastructure

10.1k

Models

Claude Haiku 4.5

Anthropic

Input tokens/M

$35

Output tokens/M

200

Context Length

Gemini 2.5 Flash

Google

$2.1

Input tokens/M

$17.5

Output tokens/M

Context Length

Gemini 2.5 Flash-Lite

Google

$0.7

Input tokens/M

$2.8

Output tokens/M

Context Length

Qianfan-Lightning

Baidu

Input tokens/M

Output tokens/M

128

Context Length

Doubao-1.5-pro-32k

Bytedance

$0.8

Input tokens/M

Output tokens/M

128

Context Length

Hunyuan-T1-20250822

Tencent

Input tokens/M

Output tokens/M

Context Length

DeepSeek-V3.1

Deepseek

Input tokens/M

$12

Output tokens/M

128

Context Length

Hunyuan-T1-latest

Tencent

Input tokens/M

Output tokens/M

Context Length

Claude Opus 4.1

Anthropic

$105

Input tokens/M

$525

Output tokens/M

200

Context Length

GLM-4.5-Flash

Chatglm

Input tokens/M

Output tokens/M

128

Context Length

GLM-4.5-AirX

Chatglm

Input tokens/M

Output tokens/M

128

Context Length

GLM-4.5-Air

Chatglm

Input tokens/M

Output tokens/M

128

Context Length

Spark X1

Iflytek

Input tokens/M

Output tokens/M

Context Length

Doubao-1.5-thinking-pro

Bytedance

Input tokens/M

$16

Output tokens/M

128

Context Length

Doubao-1.5-UI-TARS

Bytedance

$3.5

Input tokens/M

$12

Output tokens/M

128

Context Length

DeepSeek-V3

Deepseek

Input tokens/M

Output tokens/M

Context Length

Qwen3-14B

Alibaba

Input tokens/M

Output tokens/M

Context Length

ERNIE X1 Turbo

Baidu

Input tokens/M

Output tokens/M

Context Length

Gemma 3n E2B Instructed

Google

Input tokens/M

Output tokens/M

Context Length

Hunyuan-A13B

Tencent

$0.5

Input tokens/M

Output tokens/M

224

Context Length

Empowering the future, your artificial intelligence solution think tank

English 简体中文繁體中文にほんご

FirendLinks:

AI Newsletters AI Tools MCP Servers AI News AI Marketing LLM Leaderboard AI Ranking

Business Cooperation Site Map